Claude's Constitution
AI Ethics and the Lacanian Fantasy of Order
Over the past several months I’ve been working out more of my thoughts on games, virtue, and psychoanalysis. The essays have touched on rules, limits, and the uneasy relationship between coherence and desire. In a game, the rules are always incomplete and never tell the full story. They create and frame the field of play, but something is always left out and exceeds them. In the virtue essays, I argued that moral language is meaningful only insofar as it contends with its inherent and necessary limitation. Virtue becomes dangerous when it cultivates the fantasy of wholeness rather than a posture of lack.
The Myth of AI Moral Agency
AI is not a board game, and it is certainly not a moral agent. However, with the recent publication of a “constitution” by the Anthropic team—alongside its ongoing ethical dispute with the Department of Defense—something familiar is happening. We are watching an effort to stabilize authority in a space where its ground no longer feels secure.
Photo by Anthony Garand on Unsplash
AI systems have always had guardrails, policies, and review processes. This is not the birth of governance. Companies have long written documents to regulate risk. What feels different in this moment is the rhetoric. Policy becomes constitution. Compliance becomes moral language. Governance becomes character.
This is no longer simply about what the system is permitted to do. It is about what kind of “agent” the system is imagined to be; whether it can be described as possessing judgment, priorities, or even something like virtue.
Words as Master Signifiers
The significance lies less in the specific clauses than in the very act of writing a constitution at all. Constitutions are written when authority must be declared rather than assumed. They appear when legitimacy feels unstable. They declare priorities, install hierarchies, and name what will trump what.
In Lacanian psychoanalytic theory, we could call these master signifiers. They are the words that attempt to hold an unstable field in place. Language is slippery; every word requires another word to explain it. At some point, something has to stop the drift. Master signifiers are the terms that seem to anchor everything else—words that feel foundational because they promise clarity.
In Claude’s constitution, these are words like safety, ethics, corrigibility, and helpfulness. They function as organizing anchors. They tell us what outranks what. They reassure us that priorities have been named.
And yet, like any student in a philosophy seminar eventually discovers, these words never quite close the loop. “Safety” sounds solid until someone asks, “safe for whom?” “Ethics” sounds obvious until we ask, “according to which framework?” These questions are not pedantic. They reveal that the anchoring words do not eliminate ambiguity. They manage it.
For example, an AI model is asked to generate analysis for military strategy. Is refusing the request an expression of “safety”? Or does assisting in defensive planning count as “helpfulness”? What if the same system is used in one context for national security and in another for suppression? The word “ethics” does not resolve the conflict; it organizes it. It tells us which value outranks another, but it does not eliminate the tension between them. The ambiguity does not disappear. It is structured.
The Illusion of Coherence
This tension becomes sharper when we place it alongside how virtue functions. In the earlier essays, I suggested that virtue is useful precisely because it does not promise final coherence. No subject is fully integrated or complete. The moral life is not the achievement of wholeness but an ongoing reorientation toward the good. Virtue matters because it acknowledges its own limits.
When a constitution speaks of judgment, character, and values in the context of AI governance, a similar structure appears—but with a different risk. The language of hierarchy and law can begin to suggest that instability itself can be resolved through proper ordering. Instead of saying, “this system is structurally unpredictable,” the document reframes unpredictability as something that can be supervised and corrected. Uncertainty becomes oversight. Drift becomes corrigibility.
This does not make the document insincere. Oversight and corrigibility are meaningful design commitments. The concern is subtler. Moral language can quietly cultivate the impression that coherence has been secured and that ambiguity has been domesticated through hierarchy. In that sense, the constitution both reveals instability and attempts to contain it.
The very existence of a constitution signals this tension. It anticipates misuse, value conflict, and failure. It names priorities in advance of inevitable collisions. Rather than eliminating instability, it stages a response to it. If the system were fully coherent and self-evident, no constitution would be necessary.
Claude’s constitution is not unique in this effort. OpenAI’s Model Spec documents perform a similar function, though in a different register. Each lab adopts its own style—some technical, some procedural, some explicitly moral—but all publicly articulate the principles meant to guide systems whose behavior cannot be fully exhausted in advance.
These documents attempt to secure coherence in a domain that resists it. In doing so, they make visible the very instability they seek to manage.
The desire to stabilize is understandable. Systems that generate language at scale will inevitably encounter moral conflict and ambiguity. Governance documents create shared expectations about how those tensions will be handled. But beneath these efforts lies a deeper confusion: we often speak about these systems as if coherence or incoherence were properties they experience rather than effects we interpret.
AI as a Moral Mirror
AI is not a subject. It does not experience lack, desire, or ambiguity. Large language models operate through statistical probability, translating words into numerical patterns and back again. Whatever coherence or instability appears in their output is not lived from within; it is attributed by us.
In that sense, AI functions less as an autonomous moral agent and more as a mirror. It reflects patterns in language and value back to us without inhabiting them. What we experience as drift in the machine is not something alien. It echoes the way meaning itself never quite settles for us.
The question, then, is not whether we can finally render AI coherent and without lack. The question is what position we occupy when we turn to it for coherence. What do we imagine it to be? What do we ask it to do for us? How are we being formed in the process?
In previous essays on AI, I have suggested that these systems function less as solutions to lack and more as sites where our own lack becomes visible. In the next essay, I want to explore what happens when we address AI as the “subject supposed to know” and what that position reveals about us.
📚 Works Cited & Suggested Reading
Anthropic. “Claude’s Constitution.” 2026. https://www.anthropic.com/constitution.
Anthropic. Responsible Scaling Policy. 2026. https://www.anthropic.com/responsible-scaling-policy.
Fink, Bruce. The Lacanian Subject: Between Language and Jouissance. Princeton: Princeton University Press, 1995.
Lacan, Jacques. Écrits. Translated by Bruce Fink. New York: W.W. Norton, 2006.
Lacan, Jacques. The Seminar of Jacques Lacan, Book XI: The Four Fundamental Concepts of Psychoanalysis. Translated by Alan Sheridan. New York: W.W. Norton, 1981.
McGowan, Todd. Capitalism and Desire: The Psychic Cost of Free Markets. New York: Columbia University Press, 2016.
OpenAI. Model Spec: Principles and Operational Guidelines for ChatGPT. OpenAI, 2025. https://model-spec.openai.com/2025-12-18.html
Robins-Early, Nick. “Anthropic Says It ‘Cannot in Good Conscience’ Allow Pentagon to Remove AI Safeguards from Claude.” The Guardian, February 26, 2026.
Verhaeghe, Paul. On Being Normal and Other Disorders: A Clinical Introduction to Lacanian Psychoanalysis. New York: Other Press, 2004.